Optimizations in Iris Recognition A Dissertation Submitted to the Graduate School of the University of Notre Dame in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy in Computer

by Xiaomei Liu Biometric verification systems employing images of the iris are claimed to be extremely accurate, yielding no false accepts at any reasonable false reject rate. However, there are few if any large scale experimental evaluations on public iris datasets reported in the literature. We have collected an iris image dataset of over 25,000 iris images from over 300 persons (over 600 irises). When collecting the dataset, we intentionally allowed a broader image quality range than that allowed by default in current commercial iris recognition systems. The iris images used in our experiments have been, or will be, released as part of the Iris Challenge Evaluation (ICE). We reimplemented in C an open source iris recognition system, which was originally implemented in MATLAB by Libor Masek. The ICE baseline is a C++ translation of our C re-implementation with modifications for optimization in speed and memory usage. We evaluated the effects of iris image quality by using the ICE baseline system on our iris dataset. We have implemented an improved iris segmentation and eyelid detection stage compared to the ICE baseline code, and experimentally verified an improvement in both the verification and identification contexts. Replacing the ICE baseline segmentation with our improved segmentation algorithm, and keeping other modules of the ICE baseline the same, leads to an increase of over 6% in the rank-one Xiaomei Liu recognition rate and a decrease of over 4% in the equal error rate. We utilized an active contour model to refine the noise detection results and optimized the matching stage to compensate for the possible inaccuracy in iris segmentation and noise detection, which leads to another 0.95% increase in the rank one recognition rate and 0.85% decrease in the equal error rate. This research demonstrates that a more accurate iris segmentation helps to improve the overall system performance, and that the inaccuracy of iris segmen-tation and noise detection could be partly compensated for with optimizations in the matching stage.